ChatGPT is more extroverted, conscientious, and more open to experiences than the average working age human, new research from Sapia.ai has found.
If these generative AI models were job candidates responding to interview questions, what kind of personality would they project in their natural form? ChatGPT (based on GPT-3.5) and more recently GPT-4 are built using a process that includes “reinforcement learning from human feedback (RLHF)” to produce responses which are less likely to make up facts, be toxic or harmful in their sentiment compared to the earlier GPT-2. Could this friendliness and politeness be verified through the personality these models project? These were some of the questions that the team at Sapia Labs, the innovation arm of Sapia.ai, set out to answer.
“We all know ChatGPT can be prompted to respond in different ways and, as AI models, personality is not directly applicable to them. However, given these models generate responses similar to humans, we wanted to better understand the personality projected by these models when they are not prompted to be a certain personality, or in its natural form, and be able to distinguish its responses from that of a human,” Sapia.ai’s Chief Data Scientist Dr. Buddhi Jayatilleke said.
The study, a first of its kind, analysed over 6,000 responses from GPT-2, ChatGPT (GPT-3.5) and GPT-4, and cross-examined them with Sapia.ai’s dataset (currently at 4 million candidates across 47 countries with over 1.5 billion words shared by job candidates). Sapia Labs used their industry-leading personality inference models based on InterviewBERT, a fine tuned version of Google’s BERT large language model to infer the personality dimensions similar to the well known HEXACO model of personality.
Their findings provide fascinating insights into the nature of these generative models. Both ChatGPT and GPT-4 scored significantly higher on the dimensions of honesty/humility, agreeableness, and consciousness compared to GPT-2. These results align well with OpenAI’s description of how ChatGPT models differ from earlier GPT models. The higher honesty/humility and agreeableness is consistent with modesty, politeness, and friendly responses. Additionally, the greater likelihood to follow directions and provide accurate information aligns with ChatGPT’s higher conscientiousness. But the most interesting finding is that both ChatGPT and GPT-4 showed significantly higher levels of extroversion, conscientiousness, and open to experience compared to the human benchmark dataset. In other words the newer GPT models trained with human-in-the-loop project a more sociable, open minded and diligent personality.
Global brands including Joe & the Juice, Starbucks, Woolworths Group, and Qantas trust Sapia.ai to accelerate and enhance their recruitment and promotion processes, bringing them the people who belong with their brands by giving every candidate a chance to interview over a low pressure chat.